Prompt Generator for Stable Diffusion#

Note

Install ekorpkit package first.

Set logging level to Warning, if you don’t want to see verbose logging.

If you run this notebook in Colab, set Hardware accelerator to GPU.

%pip install -U --pre ekorpkit[art]

exit()

Preparing the environment#

%config InlineBackend.figure_format='retina'
%load_ext autotime

from ekorpkit import eKonf

eKonf.setLogger("WARNING")
eKonf.set_cuda(device=0)
print("version:", eKonf.__version__)

is_colab = eKonf.is_colab()
print("is colab?", is_colab)
if is_colab:
    eKonf.mount_google_drive()
project_dir = eKonf.set_workspace(workspace="/content/drive/MyDrive/workspace/", project="ekorpkit-book")
print("project_dir:", project_dir)

INFO:ekorpkit.base:Setting cuda device to ['A100-SXM4-40GB (id:0)']
INFO:ekorpkit.base:Setting EKORPKIT_WORKSPACE_ROOT to /content/drive/MyDrive/workspace/
INFO:ekorpkit.base:Setting EKORPKIT_PROJECT to ekorpkit-book
INFO:ekorpkit.base:Loaded .env from /workspace/projects/ekorpkit-book/config/.env

version: 0.1.40.post0.dev9
is colab? False
project_dir: /content/drive/MyDrive/workspace/projects/ekorpkit-book
time: 3.43 s (started: 2022-11-10 09:47:29 +00:00)

Load a Generator and Generate Prompts#

To download a certain dataset or model checkpoint, you may need to provide a HuggingFace API token. You can get one from here.


import os

# Set HuggingFace API token
os.environ["HF_USER_ACCESS_TOKEN"] = "YOUR_TOKEN"

from ekorpkit.models.art.prompt import PromptGenerator

pgen = PromptGenerator(verbose=False)

2022-11-10 09:47:35.832143: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 FMA
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.

time: 6.51 s (started: 2022-11-10 09:47:34 +00:00)

Loading a model#

Before loading a model, you need to train a model first. To train a model, refer to the Training a Generator section.

# ekorpkit/stable-prompts
pgen.load_model(model_name="ekorpkit/stable-prompts")

time: 4.97 s (started: 2022-11-10 07:43:28 +00:00)

Generating prompts#

You can generate prompts using the generate_prompts function. The function takes the following arguments:

  • prompt: The prompt to be used for generating prompts. If None, the prompt will be generated automatically.

  • num_prompts_to_generate: The number of prompts to be generated.

  • generate_images: Whether to generate images for the prompts.

  • num_samples: The number of images to be generated for each prompt.

  • For other arguments, refer to the following code.

eKonf.print(pgen.generate_config)

{'check_start_token': True,
 'max_prompt_length': 50,
 'min_prompt_length': 30,
 'num_prompts_to_generate': 10,
 'prompt': None,
 'temperature': 1.2,
 'top_k': 70,
 'top_p': 0.9}
time: 1.18 ms (started: 2022-11-10 07:43:34 +00:00)

prompts = pgen.generate_prompts(
    prompt="people looking out a lonely city street",
    num_prompts_to_generate=3,
    generate_images=True,
    num_samples=2,
)

Prompt: people looking out a lonely city street in the snow, concept art portrait of a blonde anime boy with blond hair, high detail, digital art, by greg rutkowski
../../../_images/prompt-generator_12_1.png 
Prompt: people looking out a lonely city street, artgerm and greg rutkowski and alphonse mucha, a single glowing white male wolf with blue fur, fur, wolf face, ultra realistic, concept art, intricate details
../../../_images/prompt-generator_12_3.png 
Prompt: people looking out a lonely city street, trending on artstation by Artgerm and Greg Rutkowski and Alphonse Mucha, 8k, HD
../../../_images/prompt-generator_12_5.png 
time: 1min 6s (started: 2022-11-10 07:49:23 +00:00)

Generating images for prompts#

results = pgen.generate_images(
    prompts=prompts,
    num_samples=3,
)

Prompt: people looking out a lonely city street corner. cityscape. By Greg Rutkowski, Ilya Kuvshinov, WLOP, Stanley Artgerm Lau, Ruan Jia and Fenghua Zhong, trending on ArtStation
../../../_images/prompt-generator_14_1.png 
Prompt: people looking out a lonely city street below a lake on a white background, artstation, concept art, matte painting by craig mullins and Anato Finnstark, high detail, volumetric lighting, octane render, 4K resolution
../../../_images/prompt-generator_14_3.png 
Prompt: people looking out a lonely city street at sunset, fantasy art by greg rutkowski and craig mullins, trending on artstation
../../../_images/prompt-generator_14_5.png 
time: 1min 36s (started: 2022-11-10 07:46:05 +00:00)

Generating images for one prompt#

results = pgen.imagine(
    text_prompts=prompts[0],
    num_samples=6,
    num_inference_steps=150,
    guidance_scale=10,
)

Prompt: people looking out a lonely city street in the snow, concept art portrait of a blonde anime boy with blond hair, high detail, digital art, by greg rutkowski
../../../_images/prompt-generator_16_1.png 
time: 1min 30s (started: 2022-11-10 07:51:57 +00:00)

Training a Generator#

Preparing a dataset#

You can use any dataset you want. However, the dataset should be in the format of HuggingFace Datasets.

  1. Using a dataset from HuggingFace Hub

dataset = pgen.load_dataset("Gustavosta/Stable-Diffusion-Prompts")
dataset

WARNING:datasets.builder:Using custom data configuration Gustavosta--Stable-Diffusion-Prompts-d22aeec0ba2a9fdb
WARNING:datasets.builder:Reusing dataset parquet (/content/drive/MyDrive/workspace/projects/ekorpkit-book/disco-imagen/cache/Gustavosta___parquet/Gustavosta--Stable-Diffusion-Prompts-d22aeec0ba2a9fdb/0.0.0/2a3b91fbd88a2c90d1dbbb32b460cf621d31bd5b05b934492fdef7d8d6f236ec)

DatasetDict({
    train: Dataset({
        features: ['Prompt'],
        num_rows: 73718
    })
    test: Dataset({
        features: ['Prompt'],
        num_rows: 8192
    })
})

time: 4.75 s (started: 2022-11-10 09:47:44 +00:00)

pgen.train_config.shuffle_prompts = False
pgen.trainer_config.num_train_epochs = 1
pgen.trainer_config.logging_steps = 1000
model_name = "ekorpkit/stable-prompts"

pgen.train(
    model_name=model_name,
    dataset=dataset,
    prompt_column="Prompt",
    split="train",
)

PyTorch: setting up devices
loading configuration file https://huggingface.co/distilgpt2/resolve/main/config.json from cache at /content/drive/MyDrive/workspace/projects/ekorpkit-book/disco-imagen/cache/f985248d2791fcff97732e4ee263617adec1edb5429a2b8421734c6d14e39bee.422318838d1ec4e061efb4ea29671cb2a044e244dc69229682bebd7cacc81631
Model config GPT2Config {
  "_name_or_path": "distilgpt2",
  "_num_labels": 1,
  "activation_function": "gelu_new",
  "architectures": [
    "GPT2LMHeadModel"
  ],
  "attn_pdrop": 0.1,
  "bos_token_id": 50256,
  "embd_pdrop": 0.1,
  "eos_token_id": 50256,
  "id2label": {
    "0": "LABEL_0"
  },
  "initializer_range": 0.02,
  "label2id": {
    "LABEL_0": 0
  },
  "layer_norm_epsilon": 1e-05,
  "model_type": "gpt2",
  "n_ctx": 1024,
  "n_embd": 768,
  "n_head": 12,
  "n_inner": null,
  "n_layer": 6,
  "n_positions": 1024,
  "reorder_and_upcast_attn": false,
  "resid_pdrop": 0.1,
  "scale_attn_by_inverse_layer_idx": false,
  "scale_attn_weights": true,
  "summary_activation": null,
  "summary_first_dropout": 0.1,
  "summary_proj_to_labels": true,
  "summary_type": "cls_index",
  "summary_use_proj": true,
  "task_specific_params": {
    "text-generation": {
      "do_sample": true,
      "max_length": 50
    }
  },
  "transformers_version": "4.21.2",
  "use_cache": true,
  "vocab_size": 50257
}

loading weights file https://huggingface.co/distilgpt2/resolve/main/pytorch_model.bin from cache at /content/drive/MyDrive/workspace/projects/ekorpkit-book/disco-imagen/cache/43a212e83e76bcb07f45be584cf100676bdbbbe9c13f9e5c1c050049143a832f.a83d881ec4d624fd4b5826dd026e315246c48c67504ff91c0500570e291a54ba
All model checkpoint weights were used when initializing GPT2LMHeadModel.

All the weights of GPT2LMHeadModel were initialized from the model checkpoint at distilgpt2.
If your task is similar to the task the model of the checkpoint was trained on, you can already use GPT2LMHeadModel for predictions without further training.
***** Running training *****
  Num examples = 73718
  Num Epochs = 1
  Instantaneous batch size per device = 1
  Total train batch size (w. parallel, distributed & accumulation) = 8
  Gradient Accumulation steps = 1
  Total optimization steps = 9215
Automatic Weights & Biases logging enabled, to disable set os.environ["WANDB_DISABLED"] = "true"
[9215/9215 18:54, Epoch 1/1]
Step Training Loss
1000 3.398000
2000 2.720500
3000 2.546200
4000 2.450300
5000 2.351200
6000 2.323600
7000 2.261700
8000 2.247700
9000 2.230900 

Training completed. Do not forget to share your model on huggingface.co/models =)
Waiting for W&B process to finish... (success).

Run history:


train/epoch▁▁▂▂▃▄▅▅▆▇██
train/global_step▁▁▂▂▃▄▅▆▆▇██
train/learning_rate██▇▇▆▅▄▃▃▂▁
train/loss█▄▄▂▂▂▁▁▁▁▁
train/total_flos▁
train/train_loss▁
train/train_runtime▁
train/train_samples_per_second▁
train/train_steps_per_second▁

Run summary:


train/epoch1.0
train/global_step9215
train/learning_rate0.0
train/loss2.2309
train/total_flos1.9262273860141056e+16
train/train_loss2.49658
train/train_runtime1138.5106
train/train_samples_per_second64.75
train/train_steps_per_second8.094

Synced ekorpkit/stable-prompts: https://wandb.ai/entelecheia/ekorpkit-book-prompt-generator/runs/3ig7uufw
Synced 5 W&B file(s), 0 media file(s), 0 artifact file(s) and 0 other file(s)
Find logs at: ./wandb/run-20221110_094831-3ig7uufw/logs
Configuration saved in /content/drive/MyDrive/workspace/projects/ekorpkit-book/disco-imagen/models/ekorpkit/stable-prompts/config.json
Model weights saved in /content/drive/MyDrive/workspace/projects/ekorpkit-book/disco-imagen/models/ekorpkit/stable-prompts/pytorch_model.bin

time: 19min 34s (started: 2022-11-10 09:51:17 +00:00)

pgen.load_model(model_name=model_name)
prompts = pgen.generate_prompts(
    prompt="people looking out a lonely city street",
    num_prompts_to_generate=2,
    generate_images=True,
    num_samples=2,
)

Prompt: people looking out a lonely city street, dark, dystopian, concept art, ambient lighting concept art ilya kuvshinov, arthur mores, pixar and ethereal, atmospheric, colorful, trending on artstation
../../../_images/prompt-generator_21_1.png 
Prompt: people looking out a lonely city street, concept art by greg rutkowskirein and jim kehan and greg rutkowski and zeremy lipkin and artgerm and luis royo hyper detailed,
../../../_images/prompt-generator_21_3.png 
time: 1min 14s (started: 2022-11-10 10:11:23 +00:00)

  1. Using a dataset from a text file

prompt_uri = "https://raw.githubusercontent.com/entelecheia/ekorpkit-book/main/assets/data/prompt_parrot.txt"
dataset = pgen.load_dataset("text", data_files=prompt_uri)
dataset

WARNING:datasets.builder:Using custom data configuration default-681f997af4470be8
WARNING:datasets.builder:Reusing dataset text (/content/drive/MyDrive/workspace/projects/ekorpkit-book/disco-imagen/cache/text/default-681f997af4470be8/0.0.0/21a506d1b2b34316b1e82d0bd79066905d846e5d7e619823c0dd338d6f1fa6ad)

DatasetDict({
    train: Dataset({
        features: ['text'],
        num_rows: 185
    })
})

time: 1.29 s (started: 2022-11-10 09:41:31 +00:00)

eKonf.print(pgen.train_config)
eKonf.print(pgen.trainer_config)

{'batch_size': 10000,
 'num_shuffles': 4,
 'padding': 'max_length',
 'pretrained_model_name_or_path': 'distilgpt2',
 'prompt_column': 'Prompt',
 'shuffle_prompts': True,
 'split': 'train'}
{'logging_steps': 200,
 'num_train_epochs': 1,
 'overwrite_output_dir': True,
 'per_device_train_batch_size': 1,
 'prediction_loss_only': True,
 'report_to': 'wandb',
 'run_name': 'ekorpkit/stable-prompts',
 'save_steps': 0,
 'seed': None}
time: 2.66 ms (started: 2022-11-10 09:41:37 +00:00)

pgen.train_config.shuffle_prompts = False
pgen.trainer_config.num_train_epochs = 3
model_name="ekorpkit/prompt_parrot"

pgen.train(model_name=model_name, dataset=dataset, prompt_column="text")

***** Running training *****
  Num examples = 185
  Num Epochs = 3
  Instantaneous batch size per device = 1
  Total train batch size (w. parallel, distributed & accumulation) = 8
  Gradient Accumulation steps = 1
  Total optimization steps = 72
Automatic Weights & Biases logging enabled, to disable set os.environ["WANDB_DISABLED"] = "true"
wandb: Currently logged in as: entelecheia. Use `wandb login --relogin` to force relogin
wandb version 0.13.5 is available! To upgrade, please run: $ pip install wandb --upgrade
Tracking run with wandb version 0.13.3
Run data is saved locally in /workspace/projects/ekorpkit-book/ekorpkit-book/docs/lectures/aiart/wandb/run-20221110_094236-1qbfarkj
Syncing run ekorpkit/prompt_parrot to Weights & Biases (docs)
[72/72 00:08, Epoch 3/3]
Step Training Loss 

Training completed. Do not forget to share your model on huggingface.co/models =)
Waiting for W&B process to finish... (success).

Run history:


train/epoch▁
train/global_step▁
train/total_flos▁
train/train_loss▁
train/train_runtime▁
train/train_samples_per_second▁
train/train_steps_per_second▁

Run summary:


train/epoch3.0
train/global_step72
train/total_flos145019696578560.0
train/train_loss8.11367
train/train_runtime67.9757
train/train_samples_per_second8.165
train/train_steps_per_second1.059

Synced ekorpkit/prompt_parrot: https://wandb.ai/entelecheia/ekorpkit-book-prompt-generator/runs/1qbfarkj
Synced 5 W&B file(s), 0 media file(s), 0 artifact file(s) and 0 other file(s)
Find logs at: ./wandb/run-20221110_094236-1qbfarkj/logs
Configuration saved in /content/drive/MyDrive/workspace/projects/ekorpkit-book/disco-imagen/models/ekorpkit/prompt_parrot/config.json
Model weights saved in /content/drive/MyDrive/workspace/projects/ekorpkit-book/disco-imagen/models/ekorpkit/prompt_parrot/pytorch_model.bin

time: 1min 32s (started: 2022-11-10 09:42:17 +00:00)

pgen.load_model(model_name=model_name)
prompts = pgen.generate_prompts(
    prompt="people looking out a lonely city street",
    num_prompts_to_generate=2,
    generate_images=True,
    num_samples=2,
)

Prompt: people looking out a lonely city street. vibrant nightscape landscape by Jean Eichler & Tom Riezzler, paintingstation paintingine de Jean du Fontaine
../../../_images/prompt-generator_26_1.png 
Prompt: people looking out a lonely city street street in a beautiful landscape by Steven Adams, lushly, and landscapes by Robert F. Bennett, The Lost and Skyward, a matte art by Alex Turner, matte painterabby, pixelism
../../../_images/prompt-generator_26_3.png 
time: 1min 14s (started: 2022-11-10 09:43:50 +00:00)

References#